Unmanned Aerial Vehicles (UAVs) demonstrated improved response times and safety in life-saving missions. This paper presents a Model Predictive Control (MPC) framework for autonomous UAV search missions in maritime environments, where the UAV must locate multiple castaways floating on the sea surface after a maritime incident. The approach uses receding horizon optimization to plan trajectories that balance two competing objectives: achieving rapid area coverage at high altitudes versus maintaining reliable target detection at lower altitudes. Target detection relies on Convolutional Neural Networks (CNNs), with detection performance characterized through field experiments that measure True Positive (TP) rates and False Positive (FP) rates across multiple flight altitudes. The MPC framework dynamically adjusts the UAV’s altitude and trajectory based on these altitude dependent detection statistics, enabling mission adaptive behavior that outperforms constant-altitude search patterns. Simulation results demonstrate improved search performance compared to conventional constant-altitude missions. Real-world flight experiments validate the practical applicability of the proposed framework and confirm its effectiveness in realistic maritime search scenarios.
Grid-based path planning and optimization remain widely used in robotics and autonomous navigation because real environments can be discretized into binary occupancy grids, allowing algorithms such as A* to guarantee a solution whenever a collision-free path exists. However, the resulting grid paths are often suboptimal in length and contain unnecessary directional changes, which increase traversal time, energy consumption, and control effort. This paper proposes Supercover-Based Corridor Reconstructed A* (SBCR-A*), a framework that combines an A* search backbone with an SBCR reconstruction pipeline to convert discrete grid paths into trajectories with improved path-length optimality while preserving collision-freedom. SBCR-A* utilizes a supercover-based line-of-sight formulation together with obstacle-boundary-guided refinement to replace stepwise grid paths with a small number of long, straight, feasible segments. On the Göttingen Maze Dataset, SBCR-A* achieves a path optimality of 107.72% relative to standard A*, which is the highest path optimality among all methods evaluated in this study, including recent learning-based planners and existing A*-based improvements. Moreover, when the SBCR-A* reconstruction pipeline is decoupled from its own search component and applied to paths produced by these A*-based variants, it delivers additional average path-length reductions in the range 1.85%-5.71% . To the best of our knowledge, no existing reconstruction method achieves higher path optimality on these benchmarks while providing equally general, planner-agnostic integration. Consequently, SBCR-A* can be regarded as a state-of-the-art reconstruction framework for improving the geometric quality and optimality of grid-based paths.
Increasing the flexibility of automated systems to enable high-mix low-volume production is a major trend in concurrent manufacturing, aiming to overcome limitations such as lengthy setup times and skilled labor requirements. Robotic welding systems, although widely adopted, remain cost-effective primarily for large batches with low variability. To cope with this constraint, there is increasing interest in collaborative robotic applications, where human operators and robots cooperate to achieve partial automation while maintaining the flexibility of manual labor. In this work, a collaborative robotic welding system was developed for a specific industrial case study: structural frames for luxury retail fit-outs. These custom-designed frames act as support for furniture elements and decorative wall claddings in high-end retail shops. As such, they feature high dimensional variability and are currently produced by manual metal active gas (MAG) welding of rectangular hollow section (RHS) profiles. To address the automation challenge, a digital model of the cell was first developed to compare configurations and select a feasible production system. A two-station configuration, utilizing a 400 A MAG welding source mounted on a cobot coupled with an external linear axis, was engineered and tested in the actual production environment. Process quality was ensured by selecting low heat input parameters for the vertical-up welding of S275JR RHS profiles, which produced joints failing in the base metal at tensile strengths within the base-material range. The replacement of the manual butt joints with vertical fillet welds eliminated the post-processing grinding step and, together with the automation of welding, yielded a measured lead-time reduction of up to 53
Autonomous Surface Vehicles (ASVs) are increasingly used in bathymetric surveys, environmental monitoring, and search and rescue missions. Recent research highlights ASV–UAV cooperation, leveraging the complementary strengths of both platforms for collaborative tasks. However, few ASVs are designed for education, creating a gap in hands-on marine robotics instruction. This paper presents AeroCat, a compact, low-cost ASV developed at CEFET-MG for undergraduate robotics education and later upgraded to support cooperative research. The modernized platform features improved hulls, an enhanced navigation system, and a landing/takeoff structure for small UAVs, enabling multi-robot collaboration. Key contributions include introducing AeroCat as a versatile educational and research platform, implementing autonomous navigation and control strategies, and demonstrating its application in simulation-based case studies on modeling, positioning control, and collaborative inspection.
This paper presents an obstacle-node-based Rapidly-exploring Random Tree (OB-RRT) algorithm for robotic arm path planning in constrained maintenance environments. The proposed method incorporates obstacle-node information derived from collision samples to guide tree expansion and improve exploration efficiency. The performance of OB-RRT is evaluated through simulations in 2D and 3D environments, with comparisons to RRT, GB-RRT, RRT-Connect, RRT*, and Informed-RRT*. The results indicate that, while optimization-based planners achieve better path optimality, they generally require higher computational cost. In contrast, OB-RRT provides a favorable trade-off between planning efficiency and path quality. Furthermore, the proposed method is validated on a 6-DoF robotic arm in a catenary maintenance scenario using a digital twin framework. The planned trajectories are successfully executed on a real robotic system, demonstrating feasibility for practical applications.
Numerous bio-heuristic algorithms have been developed, each with unique characteristics. This paper introduces a novel bionic algorithm, the Beaver Optimization Algorithm (BOA), inspired by the adaptive behaviors of beavers in their natural habitats. Beavers exhibit distinct flexibility and efficiency in feeding, reproduction, and evading predators, which the authors have observed and modeled in the BOA. The BOA is designed to address the path planning problem of mobile robots, drawing from beavers’ dynamic behaviors of beavers like den construction and food foraging in aquatic environments. The algorithm mathematically simulates these behaviors, incorporating elements such as den migration, localized foraging, and random hiding strategies. These simulated behaviors are translated into mechanism that offer feasible solutions for optimization problems. To evaluate the BOA’s efficacy, the authors conducted rigorous testing against 23 benchmark functions, comparing its performance with other prominent nature-inspired algorithms, including particle swarm optimization algorithms, genetic algorithms and so on. The results indicate that the BOA outperforms many of these algorithms on the majority of benchmark functions. Building on the BOA, the paper also presents a strategy for handling multi-objective optimization, leading to the development Multi-Objective Beaver Optimization Algorithm (MOBOA). This extended algorithm was tested on multi-objective benchmark functions and compared against four other multi-objective algorithms. The results demonstrate that MOBOA exhibits superior performance in these scenarios. Furthermore, the application of the BOA to mobile robot path planning is explored. The algorithm successfully optimizes path lengths ins constructed model, validating its feasibility and effectiveness through comparative analysis. Similarity, MOBOA shows promising results in addressing multi-objective path planning challenges for robots, where multiple factors such as path length, energy efficiency, and safety must be simultaneously optimized. In conclusion, the Beaver Optimization Algorithm and its multi-objective counterpart, MOBOA, represent innovative approaches in bio-heuristic optimization. Their application to complex problems like mobile robot path planning underscores their potential in providing efficient and effective solutions across various domains.
This paper addresses the output regulation (OR) bumpless transfer (BT) control problem for continuous switched linear systems (CLSs). A systematic solution procedure is developed based on mode-dependent dwell time method (MDADT) and multiple Lyapunov functions technique, which construct a kind of BT filter and BT OR equation. Firstly, the performance of output regulation bumpless transfer for CLSs is formally defined, which guarantees the continuous of control signal while asymptotic tracking for desired reference inputs and/or asymptotic rejecting undesired disturbances. Subsequently, state feedback bumpless transfer controller and error feedback bumpless transfer controller are designed to solve the OR BT problem based on MDADT switching method. A sufficient condition and an equivalent reformulation are established by multiple Lyapunov functions technique, without requiring the OR BT problem to be solvable for every individual subsystem. Finally, the effectiveness of the proposed control scheme is demonstrated through a simulation example based on an aero-engine control system.
Impelled by the evolving needs and requirements of war, industry, and society, drone technology has a long record of continuing development. A mature engineering discipline that involves artificial intelligence, biomimicry, and miniaturization has resulted from what initially was a fairly simple remote-controlled robot in early twentieth century origins, utilized primarily as a target and for reconnaissance. Drones have evolved over the years to be highly valuable high-tech warfare weapons as well as commercial, industrial, and even personal tools. Agriculture, logistics, environmental science, film production, emergency management, and many more use them increasingly. Insect-mimetic drone design-low, highly agile and biology-integrated systems that fly into regions respectful of larger drones and conduct sensitive ecological science or stealthy surveillance-has been a research subject of the past few decades. This article reviews the entire history of drone technology, its milestones, development of its multi-faceted components, including propulsion systems, navigation, sensors, communication channels, and structure, and the possibilities and limitations of the new biomimetic designs. In doing so, it attempts to provide a comprehensive understanding of the ways in which drones have evolved from primitive military innovations to some of the most sophisticated technological marvels of the time.
This Systematic Literature Review (SLR) critically analyzes the application of reinforcement learning (RL) algorithms for controlling grasp and release movements in myoelectric hand prostheses. Following Kitchenham’s methodology and incorporating PRISMA-inspired reporting principles, 72 primary studies published between 2019 and 2025 were systematically identified, evaluated, and synthesized to examine RL control strategies, algorithmic trends, reward function design, and evaluation metrics in prosthetic control systems. The findings indicate that RL has emerged as a promising approach for adaptive prosthetic control, particularly in dynamic environments characterized by noisy electromyographic (EMG) signals, variable task conditions, and user-specific motor behavior. Algorithms such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) were frequently associated with favorable reported performance due to their balance between stability, adaptability, and learning efficiency. However, the review also reveals a critical lack of standardized reward engineering and evaluation protocols, which limits reproducibility, cross-study comparability, and clinical translation. To address this gap, this review introduces taxonomy-driven analyses of reward function strategies and evaluation metrics, highlighting the predominance of generic reward formulations, heterogeneous validation settings, and limited real-world testing with prosthetic users. Overall, this work provides a structured and critical synthesis of current trends, methodological limitations, and future research directions for developing reproducible, clinically relevant, and intelligent RL-based prosthetic control systems.
In extant literature the terms automated guided vehicle (AGV), autonomous mobile robot (AMR), and mobile robot (MR) are used with different interpretations. Since the terms are closely related to technical characteristics, features, and possible applications, inconsistent use of these terms hinders their interoperability, and complicates the integration of systems across the warehousing domain. This paper defines the terms AGV, AMR, MR, based on the definitions from a systematic literature review. The study relies on English and German-language research from 2000 until 2025. Risk of bias in inclusion of studies was assessed using predefined evidence grades and content relevance grades. In total, 177 publications were included in the final analysis, clustered by locus of control, freedom of manipulation, collaborative ability, and software capability. The key findings show that the definitions of AGVs overlap with new AMR and MR concepts. The decisive factor is the ability of the vehicle or system to be autonomous. Building on the new definitions, we introduce a spectrum of autonomy maturity levels, ranging from remotely-controlled (zero autonomy) to fully autonomous (independent) systems. Understanding the nature and appropriate autonomy level of a system is essential to support scalability, reduce information asymmetries among vendors and users, and accelerate innovation in the rapidly evolving field of mobile robotics. We also identify implications for future research based on our findings in the concluding section.
Wheeled-legged robots have gained attention for their potential to combine the efficiency of wheels with the adaptability of legs. However, most prior research has focused on lightweight platforms. In this work, we present the first successful deployment of a reinforcement learning (RL) controller on a heavy wheeled-legged robot weighing over 340 kg, which, to the best of our knowledge, is one of the largest and heaviest wheeled-legged robots reported in the literature to demonstrate RL-based hybrid locomotion control. We propose a robust end-to-end control framework, incorporating a heuristic reward structure and a state estimator trained via privileged learning. Simulation experiments demonstrate accurate speed tracking performance, strong robustness to external disturbances, and graceful performance degradation under slow-response actuation. Field experiments, including plateau trials on gobi deserts, gravel, meadows, and wetlands, further demonstrate the policy’s robustness, terrain adaptability, and real-world deployability, with a maximum measured speed of 3.80 m/s achieved in outdoor tests. These simulation and field results establish a benchmark for deploying RL on heavy robotic systems under real-world constraints.
In environments with numerous automated guided vehicles, planning efficient, collision-free paths is critical. Multi-agent path finding (MAPF) addresses the challenge of planning simultaneous, collision-free paths for multiple agents. Its extension, multi-agent pickup and delivery (MAPD), integrates path planning with task assignment, where agents transport rack-type objects from pickup to delivery locations. In dense warehouse settings, a key challenge is moving the racks scheduled for shipment while treating all others as movable obstacles. We formalize this problem as multi-agent and multi-rack path finding (MARPF), which extends MAPF/MAPD by distinguishing between target and obstacle racks. Although MARPF can be formulated using integer linear programming (ILP), this approach incurs an exponential computational cost. To address this, we propose a novel multi-stage heuristic method. First, it computes a trajectory for each target rack and plans the evacuation of any obstructing racks. These rack trajectories are then decomposed into discrete tasks and assigned to agents. On grid benchmarks ranging from 7× 5 to 35× 21 with rack densities up to 90% , our method generates plans in under 10 s while maintaining success rates above 90% . For instances where ILP solutions are tractable, our approach achieves makespans within 20– 30% of the optimal, whereas ILP fails to scale to larger problems. The proposed method has the potential to improve operational efficiency in various real-world applications, including warehouse logistics, traffic management, and crowd control. The code and animations are available at https://github.com/ToyotaCRDL/MARPF-in-High-Density-Envs.
This paper describes the development of a novel instrumented wheel with onboard sensors to estimate tire deformation, together with a recurrent neural network capable of classifying terrain types. We present a functional prototype manufactured using 3D printing and a nanostructuring process that integrates piezoresistive sensors directly into the structure of an airless wheel. This process enables precise acquisition of wheel–terrain interaction signals that are informative for terrain recognition. Using these signals, along with the robot’s linear velocity, we train a bidirectional GRU to distinguish concrete, gravel, and grass terrains. Experiments on a service ground robot traversing a 180 m circuit yield a test accuracy of 95.7%, with the confusion matrix indicating consistent performance across classes. The results suggest that embedded wheel sensing combined with sequence modeling can provide accurate terrain classification while reducing payload and integration complexity. The proposed sensing wheel and inference pipeline offer a compact basis for terrain-aware mobility and potential extensions such as slip detection and payload monitoring.
Space exploration for the existence of extraterrestrial life has been a topic of paramount interest. The icy moons of our solar system are among the promising candidates. Characterized by a thick outer ice shell that covers a subsurface ocean layer, the definitive establishment of life on these moons would require examining the water physically. A comparatively convenient way to access the water is to melt through the ice shell, i.e., thermal drilling using cryobots. Before a cryobot can be used on an icy moon, the Antarctic ice sheets serve as a suitable terrestrial analog testing site. Among several constraining factors, total energy and transit time are of substantial significance for these missions. A digital twin that provides a virtual framework integrating physics and data can help assess these constrained scenarios and inform decision-making. In this work, we present two innovations. Firstly, we introduce Cryotwin, a digital twin for assessing cryobot performance. We describe its concept, use cases, and applicability across different mission phases. Secondly, we introduce a new physically consistent, state-of-the-art semi-analytical model that serves as the primary component of Cryotwin and enables performance assessment. This model, for the first time, considers both global thermal and force equilibrium while representing the complex multiphysics phase-change process observed during thermal drilling, to calculate cryobot velocity and the melt channel radius. Specifically, cryobot velocity, which controls transit time, and total efficiency, which controls total energy, are considered quantities of interest for the performance. Finally, we present a case study demonstrating the potential of Cryotwin, together with the new model, to facilitate cryobot design and development and to aid in crucial decisions regarding the optimization of total energy and transit time, subject to operating conditions.
The exposure of transmission lines in close proximity to coastal or industrial environments to high levels of pollution from chemical substances and salts is a factor that can degrade the performance of electrical systems. Given this scenario, the washing of insulators is an essential measure to mitigate the adverse effects of these pollutants. However, the process is often risky considering that technicians work in close proximity to live equipment. To address this safety concern, a drone-robot for automatic cleaning of insulator chain is proposed. This equipment, adapted with a high-pressure water pump, was designed to perform the task of cleaning insulators on live lines, thus providing a superior level of safety for technical personnel. The drone is equipped with an RGBD camera and an onboard computer, enabling tracking of the target insulator for cleaning using artificial intelligence. The movement of the drone is determined by the generation of setpoints for the flight controller based on this tracking. The results were obtained from experimental tests conducted on high voltage power transmission lines, demonstrating consistent performance in terms of flight stability and cleaning efficiency. These findings provide strong evidence of the operational robustness and practical feasibility of the proposed drone-robot system in high-risk electrical environments.
Effective bimanual collaboration represents a significant research objective for Large Language Model (LLM)-driven nursing robots. However, current approaches are constrained by critical limitations: the single-thread LLM task planner (ST-Planner) lacks co-scheduling, while the conventional multi-agent framework (DABICO) suffers from inefficient information interaction between agents, consequently compromising performance in bimanual collaboration tasks. To overcome these limitations, this study introduces Role-playing Dual Agents (RoDA), a novel dual-agent collaboration framework augmented by LLM Role-Playing. This framework implements the nursing robot as an LLM-based dual-agent system wherein each agent assumes the role of either the left or right arm. Through meticulously crafted contextual prompts explicitly defining specific identity attributes and conversational protocols for each limb, these agents facilitate high-quality collaborative dialogue reflective of their designated roles. Evaluation of RoDA was conducted through four MuJoCo simulation scenarios encompassing all four categories of bimanual collaboration, diverse task sequences, and different degrees of workspace overlap. The experimental results demonstrate that the role-playing mechanism enhances dialogue normalization, accuracy, and information richness. This high-quality interaction enables superior task planning performance, with RoDA surpassing both the ST-Planner and the baseline DABICO framework. Specifically, RoDA achieved a 100% success rate across all scenarios, exhibiting an average performance improvement of 16.3% in the Step metric and 70.2% in the Restep metric relative to DABICO. Furthermore, enhancement of performance is achieved without the requirement for LLM fine-tuning, offering advantages including flexibility, immediacy, and low development cost. Finally, RoDA was demonstrated through practical experiments on a dual-arm nursing robot.
Hyper-realistic robot heads represent a significant advancement in humanoid robotics, offering new possibilities for natural human-robot interaction in healthcare, education, and social assistance. This review examines the evolution, design principles, and implementation challenges of highly anthropomorphic robot heads, complementing existing engineering surveys with psychological foundations, quantitative analysis of implementation trends across two decades, and systematic coverage of evaluation methodologies. The historical development traces from early prototypes to current state-of-the-art systems, including commercially available platforms. The mechanical design of realistic robot faces requires careful integration of numerous facial components, each presenting unique engineering challenges. Actuator technologies ranging from servo motors to pneumatic systems enable the complex movements necessary for facial expressions. Implementation of facial expressions relies on established emotion theories that map facial muscle movements to specific emotions through standardized coding systems. Beyond basic categorical emotions, more complex emotional states and their expressions are also addressed. The review examines how different facial features contribute to human-likeness and factors affecting expression recognition. Persistent challenges are addressed including the uncanny valley effect, achieving natural movement dynamics, and balancing mechanical complexity with practical functionality, as well as ethical and cultural considerations in social robotics. Current applications in research platforms, therapeutic settings, educational environments, and social assistance are highlighted, while considering future directions for this rapidly evolving field.
Loitering Munition (LM) swarms pose a growing challenge to modern air-defense systems. A promising strategy is to utilize Loyal Wingmen (LWs) as a counter-UAV system (c-UAV) in a scenario defined as Cooperative Threat Engagement with Heterogeneous Drone Swarms (CTEDS). Unlike prior work, which optimizes policies based on simplified observations of positions and velocities, this study enhances situational awareness by integrating LiDAR perception, temporal history, and shared observations from allies. To handle this complexity, we propose a two-stage pipeline that combines supervised pre-training with fine-tuning using reinforcement learning (RL). Results show that the architecture extracts meaningful policies from high-dimensional and dynamic observations. This supports the feasibility of expanding perception as a foundation for scalable cooperative autonomy in CTEDS. The approach highlights the potential of graph attention networks for cooperative air combat and points to future research on hierarchical neural policies and multi-agent reinforcement learning with communication.
Existing distance-based formation control laws often lead to flip, flex, and reflection ambiguities in a desired formation. To avoid these ambiguities, sign-distance rigidity was introduced. This paper presents two formation control laws based on sign-distance rigidity for double-integrator agents. We propose a combined flocking and signed distance-based formation control law, enabling agents to achieve velocity consensus and converge to the desired formation without ambiguities. Then, we demonstrate that the proposed control law is also applicable to address the formation tracking control problem, where the leader of the multi-agent system moves with a constant velocity. We show that the followers under this control law are able to track the leader’s velocity and achieve the desired formation without ambiguities. Next, we propose a robust adaptive signed distance-based formation control law and show that agents not only achieve the desired formation without ambiguities, but also effectively reject unknown time-varying disturbances without requiring knowledge of their upper bounds. The stability analysis of both control laws is rigorously studied, and simulation results are also provided to validate the effectiveness of the proposed control laws.
